⚡ electrical engineering

Informativity and Identifiability for Identification of Networks of Dynamical Systems

This paper demonstrates how Gröbner bases can be employed to establish sufficient conditions for informativity and to investigate generic local identifiability in networks of dynamical systems by analyzing signal spectra, transfer function ranks, and the dimension of associated fibers.

Original authors: Anders Hansson, João Victor Galvão da Mata, Martin S. Andersen

Published 2026-02-27

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Anders Hansson, João Victor Galvão da Mata, Martin S. Andersen

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a detective trying to figure out how a complex machine works, but you can't take it apart. You can only see a few of its moving parts and hear a few of its sounds. This machine is a network of dynamical systems—think of it as a city's traffic grid, a power grid, or even a social media network where everyone influences everyone else.

The paper by Hansson, Mata, and Andersen is essentially a new detective's handbook for solving these puzzles. It answers two big questions:

Do we have enough clues? (Informativity)
Can we actually solve the puzzle uniquely? (Identifiability)

Here is the breakdown using simple analogies.

1. The Setup: The Black Box City

Imagine a city where every intersection (a "node") sends traffic signals to other intersections.

The Goal: You want to map out exactly how every intersection talks to every other one (the "transfer functions").
The Problem: You can't see the whole city. You can only measure traffic at a few specific intersections (partial measurement), and some of the roads might be one-way or have traffic lights you already know about (known transfer functions).
The Noise: There's always random noise—accidents, weather, or people taking detours—that messes up your data.

2. Question One: Do We Have Enough Clues? (Informativity)

Before you can solve the mystery, you need to make sure your clues are good enough. If you only look at a quiet street at 3 AM, you won't learn much about how the city works. You need "excitement."

The Analogy: Imagine trying to figure out the layout of a dark room by throwing a ball around.
- If you throw the ball in a straight line and it hits a wall, you learn something.
- If you throw it randomly in every direction and it bounces off every corner, you learn the whole shape of the room.
The Paper's Insight: The authors say you need to "throw the ball" (send signals) in enough different directions so that every part of the network gets excited. They provide a mathematical rule (using something called Gröbner bases, which we'll explain in a moment) to check if your signals are strong and varied enough to reveal the whole network structure.
The "Graph" Trick: They also use a map (a graph) to count the number of independent paths the signals can take. If there are enough distinct paths from your "throwing spot" to your "listening spot," you have enough information.

3. Question Two: Can We Solve It Uniquely? (Identifiability)

Okay, you have good clues. But can you be sure there is only one possible map that fits those clues? Or could there be two different city layouts that look exactly the same from your limited viewpoint?

The Analogy: Imagine you are trying to guess a secret recipe by tasting the soup.
- Identifiable: If the soup tastes salty, and the only ingredient that makes it salty is salt, you know for sure salt is in there.
- Not Identifiable: If the soup tastes sweet, it could be sugar, honey, or maple syrup. You can't tell which one it is just by tasting.
The Paper's Insight: The authors developed a way to check if the "soup" (the data) has a unique "recipe" (the network structure). They treat the network as a giant algebraic equation. If the equation has only one solution (or a finite number of very specific solutions), the network is identifiable.
The Twist: Sometimes, even if you can't identify the whole city, you can identify a specific neighborhood (a sub-network). The paper shows you how to zoom in on just the part you care about, even if the rest of the city is a mystery.

4. The Secret Weapon: Gröbner Bases

You might wonder, "How do they actually do the math?"

The Analogy: Imagine you have a massive jigsaw puzzle, but the pieces are mixed up with thousands of other pieces, and some pieces are missing. Trying to solve it by hand is impossible.
The Tool: Gröbner bases are like a super-smart robot that sorts the puzzle pieces. It takes a messy pile of algebraic equations and rearranges them into a clean, organized list where the solution pops out clearly.
Why it matters: In the past, if you knew some parts of the network (like "we know this road is a highway"), the math got too messy to solve. This paper shows how to use the "robot" (Gröbner bases) to handle those known parts effortlessly, making it possible to solve much larger and more complex networks than before.

5. Real-World Examples

The paper tests this on a few scenarios:

Simple Loop: A small network where they prove you can figure out the connections if you measure the right spots.
The "Double Trouble" Case: They show a scenario where the network looks the same from two different angles. This proves that sometimes, even with good data, you might get two different answers (local identifiability vs. global identifiability).
The Big City: They tested their method on a network with 5 nodes and 11 unknown connections. A computer using their method solved it in 11 minutes. Without this new method, a human trying to do this by hand would likely give up.

Summary

This paper is a mathematical toolkit for engineers and scientists. It tells them:

Where to look: Which signals to measure to get the most information.
What to measure: How to handle situations where you can't see everything or where some parts are already known.
How to calculate: Using advanced algebra (Gröbner bases) to ensure the solution is unique and correct.

It's like upgrading from a magnifying glass to a high-tech scanner, allowing us to understand complex, interconnected systems (like power grids, biological cells, or communication networks) with much greater confidence and precision.

1. Problem Statement

The paper addresses two fundamental challenges in the system identification of networks of dynamical systems:

Informativity: Determining whether the external signals (inputs and measured outputs) contain sufficient information to uniquely determine the predictor model.
Identifiability: Determining whether the underlying network topology and open-loop transfer functions can be uniquely recovered from the identified predictor.

The authors aim to generalize existing literature by handling:

Partial Measurements: Not all node signals are measured.
Direct Feed-through: Direct paths from inputs to measured outputs.
Known Transfer Functions: Some network components may be known a priori.
Constraints: Rational constraints relating different transfer functions.
Sub-networks: Identifying specific parts of a larger network.

2. Methodology

The authors employ a framework combining Algebraic Geometry and Graph Theory to analyze these properties.

A. System Modeling

The network is modeled as a linear dynamic system using the forward shift operator $q$ :
$P(q)w_k = Q(q)u_k$
$z_k = R(q)w_k + S(q)u_k$
Where $w_k$ are node signals, $u_k$ are inputs (split into known $r_k$ and noise $e_k$ ), and $z_k$ are outputs. The system is transformed into an innovation form to define a predictor $\hat{y}_k = W(q)z_k$ .

B. Informativity Analysis

Informativity is defined as the condition where the predictor $W(q)$ is uniquely determined by the data.

Spectral Condition: A sufficient condition is that the spectrum of the external signals is positive definite and the transfer function from inputs to the predictor inputs has full generic rank.
Graph-Theoretic Approach: The authors define a directed graph where vertices represent signal nodes and edges represent non-zero transfer functions. They prove that the generic rank of the closed-loop transfer function matrix $\Pi$ equals the maximum number of vertex-disjoint paths from input nodes to output nodes.
Algebraic Geometry Approach (Gröbner Bases): To handle constraints and known parameters, the authors formulate the rank condition using ideals and Gröbner bases. By constructing an ideal generated by the minors of the transfer function matrix, they can algorithmically determine the generic rank and identify sets of parameters where rank deficiency occurs.

C. Identifiability Analysis

Identifiability is investigated via the concept of Generic Local Network Identifiability.

Fiber Dimension: The problem is mapped to finding the dimension of the "fiber" (the set of solutions $X$ ) for a given closed-loop transfer function $G_c$ .
Theorem: The network is generically locally identifiable if and only if the dimension of the parameter space ( $V_o$ ) equals the dimension of the image space ( $V_c$ ).
Computation: This is computed by determining the dimension of the affine variety defined by the rational equations relating parameters to the closed-loop transfer function, again utilizing Gröbner bases.

3. Key Contributions

Generalized Framework: The paper extends previous results to accommodate partial measurements, direct feed-through terms, and known transfer functions, which were often excluded in prior works.
Unified Algebraic-Geometric Characterization: It provides a rigorous method using Gröbner bases to handle rational constraints and known parameters seamlessly, offering a more flexible approach than purely graph-theoretic methods which often assume all non-zero entries are free variables.
Sub-network Identification: The authors derive specific transformations to analyze sub-networks. They show how to treat unmeasured parts of the network as exogenous inputs or how to measure specific linear combinations of unmeasured nodes to avoid estimating the entire network.
Counter-Example to Conjectures: Through Example 3, the paper demonstrates that generic local identifiability does not necessarily imply global identifiability in networks where transfer functions appear multiple times (symmetry), challenging a conjecture found in previous literature.

4. Key Results

Informativity Condition: A signal set is informative if the external signal spectrum is positive definite and the generic rank of the transfer matrix $\Pi$ is full. This rank is computable via maximum disjoint paths in a graph or via Gröbner bases.
Identifiability Condition: A network is generically locally identifiable if the dimension of the parameter variety equals the dimension of the closed-loop transfer function variety.
Algorithmic Verification: The authors provide a MATLAB-based workflow using Gröbner basis calculations to verify these conditions for complex networks that are intractable by hand.
Handling Constraints: The methodology successfully handles cases where transfer functions are known or constrained (e.g., $G_1 = G_2$ ), showing that these constraints reduce the dimension of the parameter space, potentially aiding identifiability.

5. Significance

Theoretical Advancement: The paper bridges the gap between system identification and algebraic geometry, providing a robust mathematical foundation for analyzing complex network structures that were previously difficult to analyze.
Practical Applicability: By introducing Gröbner bases, the authors offer a computational tool capable of handling large-scale networks and complex constraints (like known parameters), which is crucial for real-world applications where full network measurement is impossible.
Refinement of Identifiability Concepts: The distinction made between local and global identifiability, particularly in the context of repeated transfer functions, clarifies limitations in current identification theories and suggests directions for future research into the interplay between informativity and identifiability.

In summary, this paper provides a comprehensive, mathematically rigorous toolkit for determining whether a dynamical network can be identified from partial data, leveraging advanced algebraic geometry to overcome the limitations of traditional graph-theoretic approaches.